Main

In your final repo, there should be an R markdown file that organizes all computational steps for evaluating your proposed Facial Expression Recognition framework.

This file is currently a template for running evaluation experiments. You should update it according to your codes but following precisely the same structure.

Step 0 set work directories

set.seed(2020)

Provide directories for training images. Training images and Training fiducial points will be in different subfolders.

train_dir <- "../data/train_set/" # This will be modified for different data sets.
train_image_dir <- paste(train_dir, "images/", sep= "")
train_pt_dir <- paste(train_dir,  "points/", sep= "")
train_label_path <- paste(train_dir, "label.csv", sep= "")

Part 1: Baseline Model (GBM and logistic regression LASSO)

Step 1: set up controls for evaluation experiments.

In this chunk, we have a set of controls for the evaluation experiments.

(T/F) cross-validation on the training set
(T/F) reweighting the samples for training set
(number) K, the number of CV folds
(T/F) process features for training set
(T/F) run evaluation on an independent test set
(T/F) process features for test set

run.cv <- TRUE # run cross-validation on the training set
sample.reweight <- TRUE # run sample reweighting in model training
K <- 5  # number of CV folds
run.feature.train <- TRUE # process features for training set
run.test <- TRUE # run evaluation on an independent test set
run.feature.test <- TRUE # process features for test set

By Using cross-validation, we want to use different specifications to compare the performance of models. We setup t as tune parameter for GBM.

k = c(50,100,150,200,250,300)
model_labels = paste("GBM with k =", k)

Step 2: import data and train-test split

#train-test split
info <- read.csv(train_label_path)
n <- nrow(info)
n_train <- round(n*(4/5), 0)
train_idx <- sample(info$Index, n_train, replace = F)
test_idx <- setdiff(info$Index, train_idx)

If you choose to extract features from images, such as using Gabor filter, R memory will exhaust all images are read together. The solution is to repeat reading a smaller batch(e.g 100) and process them.

n_files <- length(list.files(train_image_dir))
image_list <- list()
for(i in 1:100){
   image_list[[i]] <- readImage(paste0(train_image_dir, sprintf("%04d", i), ".jpg"))
}

Fiducial points are stored in matlab format. In this step, we read them and store them in a list.

# #function to read fiducial points
# #input: index
# #output: matrix of fiducial points corresponding to the index
# readMat.matrix <- function(index){
#      return(round(readMat(paste0(train_pt_dir, sprintf("%04d", index), ".mat"))[[1]],0))
# }
# #load fiducial points
# fiducial_pt_list <- lapply(1:n_files, readMat.matrix)
# save(fiducial_pt_list, file="../output/fiducial_pt_list.RData")

load("../output/fiducial_pt_list.RData")

Step 3: construct features and responses

The follow plots show how pairwise distance between fiducial points can work as feature for facial emotion recognition.
- In the first column, 78 fiducials points of each emotion are marked in order.
- In the second column distributions of vertical distance between right pupil(1) and right brow peak(21) are shown in histograms. For example, the distance of an angry face tends to be shorter than that of a surprised face.
- The third column is the distributions of vertical distances between right mouth corner(50) and the midpoint of the upper lip(52). For example, the distance of an happy face tends to be shorter than that of a sad face.

Figure1

feature.R should be the wrapper for all your feature engineering functions and options. The function feature( ) should have options that correspond to different scenarios for your project and produces an R object that contains features and responses that are required by all the models you are going to evaluate later.

feature.R
Input: list of images or fiducial point
Output: an RData file that contains extracted features and corresponding responses

source("../lib/feature.R")
tm_feature_train <- NA
if(run.feature.train){
  tm_feature_train <- system.time(dat_train <- feature(fiducial_pt_list, train_idx))
  save(dat_train, file="../output/feature_train.RData")
}else{
  load(file="../output/feature_train.RData")
}
tm_feature_test <- NA
if(run.feature.test){
  tm_feature_test <- system.time(dat_test <- feature(fiducial_pt_list, test_idx))
  save(dat_test, file="../output/feature_test.RData")
}else{
  load(file="../output/feature_test.RData")
}

load(file="../output/feature_test.RData")
load(file="../output/feature_train.RData")

Part 2: Baseline Model (SVM)

Step 1 - 3: Same as the baseline model

Same as Baseline model

Using cross-validation, we compare the performance of models with different specifications. In the following chunk of code, we tune parameter cost_svm (number of shrinkage) for the Support vector machine (SVM)

cost_svm = c(0.00001, 0.0001, 0.001, 0.01, 0.1)
model_labels_svm = paste("SVM with cost =", cost_svm) 
model_labels_svm

## [1] "SVM with cost = 1e-05" "SVM with cost = 1e-04" "SVM with cost = 0.001"
## [4] "SVM with cost = 0.01"  "SVM with cost = 0.1"

Step 4: Train a classification model with training features and responses

source("../lib/SVM_model.R")

# SVM Cross-validation
cost = c(0.00001, 0.0001, 0.001, 0.01, 0.1)
model_labels_svm = paste("SVM with cost =", cost) 
model_labels_svm

err_svm <- matrix(0, nrow = length(cost), ncol = 2)
for(i in 1:length(cost)){
   print(paste("cost=", cost[i]))
   err_svm[i,] <- CV_SVM(dat_train, K=5, cost[i])
   save(err_svm, file="../output/err_svm.RData")
 }

#Load visualization of cross validation results of svm
load("../output/err_svm.RData")
err_svm <- as.data.frame(err_svm)
colnames(err_svm) <- c("mean_error", "sd_error")
cost = c(0.00001, 0.0001, 0.001, 0.01, 0.1)
err_svm$cost = as.factor(cost)
err_svm %>% ggplot(aes(x = cost, y = mean_error, ymin = mean_error - sd_error, ymax = mean_error + sd_error)) +
    geom_crossbar() +
    theme(axis.text.x = element_text(angle = 90, hjust = 1))

Find the best cost for SVM model

cost_best_svm <- cost[which.min(err_svm[,1])]
#par_best_svm <- list(cost = cost_best_svm)
# Training 
tm_train_svm = NA
tm_train_svm <- system.time(fit_train_svm <- svm(label ~., data = dat_train, kernel = "linear", cost = cost_best_svm) )
#Save and load model
saveRDS(fit_train_svm, "../output/fit_train_svm.RDS")

fit_train_svm <- readRDS("../output/fit_train_svm.RDS")
# Testing
tm_test_svm=NA
tm_test_svm <- system.time(pred_svm <- predict(fit_train_svm, dat_test))
# Evaluation
accu_svm <- mean(dat_test$label == pred_svm)
real_label = dat_test$label %>% as.character() %>% as.numeric()
pred_value_svm  = pred_svm %>% as.character() %>% as.numeric()
confusionMatrix(pred_svm,dat_test$label)

## Confusion Matrix and Statistics
## 
##           Reference
## Prediction   0   1
##          0 467  98
##          1   6  29
##                                          
##                Accuracy : 0.8267         
##                  95% CI : (0.794, 0.8561)
##     No Information Rate : 0.7883         
##     P-Value [Acc > NIR] : 0.01096        
##                                          
##                   Kappa : 0.2934         
##                                          
##  Mcnemar's Test P-Value : < 2e-16        
##                                          
##             Sensitivity : 0.9873         
##             Specificity : 0.2283         
##          Pos Pred Value : 0.8265         
##          Neg Pred Value : 0.8286         
##              Prevalence : 0.7883         
##          Detection Rate : 0.7783         
##    Detection Prevalence : 0.9417         
##       Balanced Accuracy : 0.6078         
##                                          
##        'Positive' Class : 0              
##

cost_choose <- cost[which.min(err_svm[,1])]
cat("The accuracy of model: cost =", cost_choose, "is", accu_svm*100, "%.\n")

## The accuracy of model: cost = 0.001 is 82.66667 %.

# The AUC for SVM model
AUC_SVM = auc_roc(real_label, pred_value_svm)
AUC_SVM

## [1] 0.8275601

Summarize Running Time

Prediction performance matters, so does the running times for constructing features and for training the model, especially when the computation resource is limited.

Model_performace <- function(time_feature_train, time_feature_test, time_train, time_test){
  cat("Time for constructing training features=", time_feature_train[1], "s \n")
  cat("Time for constructing testing features=", time_feature_test[1], "s \n")
  cat("Time for training model=", time_train[1], "s \n") 
  cat("Time for testing model=", time_test[1], "s \n")
} 
cat("The accuracy of the SVM model: cost =", cost[which.min(err_svm[,1])], "is", accu_svm*100, "%.\n")

## The accuracy of the SVM model: cost = 0.001 is 82.66667 %.

cat("The auc value for the SVM model is",AUC_SVM * 100, "%.\n")

## The auc value for the SVM model is 82.75601 %.

Model_performace(tm_feature_train, tm_feature_test, tm_train_svm, tm_test_svm)

## Time for constructing training features= 0.994 s 
## Time for constructing testing features= 0.192 s 
## Time for training model= 94.734 s 
## Time for testing model= 8.385 s

Part 2: Improved Model (SVM)

Step 1 - 2 : Same as Baseline SVM model

Step 3: construct features and responses

Feature extraction is the same as Baseline model, and we improved the feature by using Principal Components Analysis for feature selection.

feature_improved.R should be the wrapper for all your feature engineering functions and options. The function feature_improved() should have options that correspond to different scenarios for your project and produces an R object that contains features and responses that are required by all the models you are going to evaluate later.
+ feature_improved.R
+ Input: train/test data
+ Output: an RData file that contains extracted features and corresponding responses

source("../lib/feature_improved.R") # change file name
data_test_imp <- dat_test[ ,-ncol(dat_test)]
data_train_imp <- dat_train[ ,-ncol(dat_train)]

tm_feature_train_imp <- NA
if(run.feature.train){
  tm_feature_train_imp <- system.time(data_train_imp <- feature_improved(data_train_imp, index = NULL))
}

index_train_pca <- ncol(data_train_imp) 

tm_feature_test_improved <- NA
if(run.feature.test){
  tm_feature_test_imp <- system.time(data_test_imp <- feature_improved(data_test_imp, index = index_train_pca))
}

save(data_train_imp, file="../output/feature_train_imp.RData")
save(data_test_imp, file="../output/feature_test_imp.RData")

data_train_imp <- as.data.frame(data_train_imp)
data_train_imp$label <- dat_train$label
colnames(data_train_imp)

##  [1] "PC1"   "PC2"   "PC3"   "PC4"   "PC5"   "PC6"   "PC7"   "PC8"   "PC9"  
## [10] "PC10"  "PC11"  "PC12"  "PC13"  "PC14"  "PC15"  "PC16"  "PC17"  "PC18" 
## [19] "PC19"  "PC20"  "PC21"  "PC22"  "PC23"  "PC24"  "PC25"  "PC26"  "PC27" 
## [28] "PC28"  "PC29"  "PC30"  "PC31"  "label"

data_test_imp <- as.data.frame(data_test_imp)
data_test_imp$label <- dat_test$label
colnames(data_test_imp)

##  [1] "PC1"   "PC2"   "PC3"   "PC4"   "PC5"   "PC6"   "PC7"   "PC8"   "PC9"  
## [10] "PC10"  "PC11"  "PC12"  "PC13"  "PC14"  "PC15"  "PC16"  "PC17"  "PC18" 
## [19] "PC19"  "PC20"  "PC21"  "PC22"  "PC23"  "PC24"  "PC25"  "PC26"  "PC27" 
## [28] "PC28"  "PC29"  "PC30"  "PC31"  "label"

Step 4: Train a classification model with training features and responses

source("../lib/SVM_model_weighted.R")

#SVM Cross-validation
cost = c(0.00001,0.0001,0.001,0.01,0.1,1)
err_svm_imp <- matrix(0, nrow = length(cost), ncol = 2)
for(i in 1:length(cost)){
   print(paste("cost:", cost[i]))
   err_svm_imp[i,] <- CV_SVM_weight(data_train_imp, K = 5, cost[i])
   saveRDS(err_svm_imp, file="../output/err_svm_imp.RDS")
}
err_svm_imp
cost_best_svm_imp <- cost[which.min(err_svm_imp[,1])]
cost_best_svm_imp

From err_svm_imp, we obtained that the cost didn’t influence the error mean and the error’s standard deviation

cost_best_svm_imp = 0.01
# Training
tm_train_svm_imp = NA
temp <- ovun.sample(label ~ ., data = data_train_imp, method = "over", p = 0.3)$data
tm_train_svm_imp <- system.time(fit_train_svm_imp <- svm(label ~., data = temp, kernel = "linear", cost = cost_best_svm_imp) )
#Save and load model
saveRDS(fit_train_svm_imp, "../output/fit_train_svm_imp.RDS")

fit_train_svm_imp <- readRDS("../output/fit_train_svm_imp.RDS")
# Testing 
tm_test_svm_imp = NA
tm_test_svm_imp <- system.time(pred_svm_imp <- predict(fit_train_svm_imp, data_test_imp))

# Evaluation
accu_svm_imp <- mean(dat_test$label == pred_svm_imp)
accu_svm_imp

## [1] 0.7616667

pred_value_svm_imp  = pred_svm_imp %>% as.character() %>% as.numeric()
real_label = dat_test$label %>% as.character() %>% as.numeric()

confusionMatrix(pred_svm_imp, dat_test$label)

## Confusion Matrix and Statistics
## 
##           Reference
## Prediction   0   1
##          0 446 116
##          1  27  11
##                                           
##                Accuracy : 0.7617          
##                  95% CI : (0.7255, 0.7952)
##     No Information Rate : 0.7883          
##     P-Value [Acc > NIR] : 0.9488          
##                                           
##                   Kappa : 0.0397          
##                                           
##  Mcnemar's Test P-Value : 1.854e-13       
##                                           
##             Sensitivity : 0.94292         
##             Specificity : 0.08661         
##          Pos Pred Value : 0.79359         
##          Neg Pred Value : 0.28947         
##              Prevalence : 0.78833         
##          Detection Rate : 0.74333         
##    Detection Prevalence : 0.93667         
##       Balanced Accuracy : 0.51477         
##                                           
##        'Positive' Class : 0               
##

print(paste("The accuracy of model: cost =", cost_best_svm_imp, "is", accu_svm_imp * 100, "%"))

## [1] "The accuracy of model: cost = 0.01 is 76.1666666666667 %"

# The AUC for SVM model
AUC_SVM_imp = auc_roc(real_label, pred_value_svm_imp)
AUC_SVM_imp

## [1] 0.541534

Summarize Running Time

Prediction performance matters, so does the running times for constructing features and for training the model, especially when the computation resource is limited.

print(paste("The accuracy of the SVM improved model: cost =", cost_best_svm_imp , "is", accu_svm_imp * 100, "%"))

## [1] "The accuracy of the SVM improved model: cost = 0.01 is 76.1666666666667 %"

print(paste("The auc value for tje SVM improved model is", AUC_SVM_imp * 100, "%"))

## [1] "The auc value for tje SVM improved model is 54.1533995130174 %"

Model_performace(tm_feature_train_imp, tm_feature_test_imp, tm_train_svm_imp, tm_test_svm_imp)

## Time for constructing training features= 170.779 s 
## Time for constructing testing features= 9.538 s 
## Time for training model= 0.719 s 
## Time for testing model= 0.054 s

Part 3: Appendix – Additional Models Tried (Random Forest, VGG16, XGboost, K Nearest Neighbors (KNN))

VGG16

Random Forest

XGboost

K Nearest Neighbors (KNN)

###Reference - Du, S., Tao, Y., & Martinez, A. M. (2014). Compound facial expressions of emotion. Proceedings of the National Academy of Sciences, 111(15), E1454-E1462.

Main

Tianle Zhu

Step 0 set work directories

Part 1: Baseline Model (GBM and logistic regression LASSO)

Step 1: set up controls for evaluation experiments.

Step 2: import data and train-test split

Step 3: construct features and responses

Part 2: Baseline Model (SVM)

Step 1 - 3: Same as the baseline model

Step 4: Train a classification model with training features and responses

Summarize Running Time

Part 2: Improved Model (SVM)

Step 1 - 2 : Same as Baseline SVM model

Step 3: construct features and responses

Step 4: Train a classification model with training features and responses

Summarize Running Time

Part 3: Appendix – Additional Models Tried (Random Forest, VGG16, XGboost, K Nearest Neighbors (KNN))

VGG16

Random Forest

XGboost

K Nearest Neighbors (KNN)